When modifying a code recently, when using the network for reasoning, it was found that changing the batch size size of the test set each time would result in different reasoning results and even erroneous results. Later, it was found that BN layer was defined in the network. BN layer would transform the data in a Batch into positive distribution during the training process. During the reasoning process, the data were processed using the parameters in the training process.
Problem Background: This problem is encountered in climbing a shopping mall on a certain evening. The original plan was to use Selenium+Chrome Driver+Mitm Proxy for happy brushing. However, after a few days, it was found that the brushing could not come out, and it would jump directly to the login interface (obviously, it was encountered with reverse climbing) To tell the truth, this is the first time that selenium has been used to climb backwards, so large-scale tests and comparisons have been made.
For a small white, there is a long way to go from understanding Batch Normalization (hereinafter referred to as BN) to using BN correctly.Make a record here.
The most searched thing about BN on the Internet is the derivation of principles and the source of relevant papers.
But this does not help us in actual use, and it is not helpful to partners who need to use it quickly.
Step 1: first write a simple python program print.py, the code is as follows: print "call success." 执行结果： >>> ================================ RESTART ================================ >>> call success. >>> 步骤2：使用bat批处理程序执行python程序，“callpython.bat